Subject: Re: [PATCH 1/2] usb: host: Implement workaround for Erratum A-009611
To: yinbo.zhu@nxp.com
References: <20170911094132.38456-1-yinbo.zhu@nxp.com>
Cc: Mathias Nyman <mathias.nyman@intel.com>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        "open list:USB XHCI DRIVER" <linux-usb@vger.kernel.org>,
        open list <linux-kernel@vger.kernel.org>
From: Mathias Nyman <mathias.nyman@linux.intel.com>
Message-ID: <59BA73B7.2080509@linux.intel.com>
Date: Thu, 14 Sep 2017 15:19:03 +0300
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101
 Thunderbird/38.8.0
MIME-Version: 1.0
In-Reply-To: <20170911094132.38456-1-yinbo.zhu@nxp.com>
Content-Type: text/plain; charset=windows-1252; format=flowed
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1781
Lines: 37

On 11.09.2017 12:41, yinbo.zhu@nxp.com wrote:
> From: "yinbo.zhu" <yinbo.zhu@nxp.com>
>
> Description: This is a occasional problem where the software
> issues an End Transfer command while a USB transfer is in progress,
> resulting in the TxFIFO  being flushed when the lower layer is waiting
> for data,causing the super speed (SS) transmit to get blocked.
> If the End Transfer command is issued on an IN endpoint to
> flush out the pending transfers when the same IN endpoint
> is doing transfers on the USB, then depending upon the timing
> of the End Transfer (and the resulting internal FIFO flush),the
> lower layer (U3PTL/U3MAC) could get stuck waiting for data
> indefinitely. This blocks the transmission path on the SS, and no
> DP/ACK/ERDY/DEVNOTIF packets can be sent from the device.
> Impact: If this issue happens and the transmission gets blocked,
> then the USB host aborts and resets/re-enumerates the device.
> This unblocks the transmitt engine and the device functions normally.
>
> Workaround: Software must wait for all existing TRBs to complete before
> issuing End transfer command.

This doesn't seem like the correct way to solve the issue.
This workaround will prevent canceling URBs.

we need to stop the endpoint ring before we can touch the TRBs on the ring.
If we want to cancel a URB we need to either hop over its TRBs on the ring,
or turn them to no-ops. The canceled URB is only given back after this.

Imagine a USB camera queuing ~100 TRBs in advance, then user wants to change video
mode and cancel the pending URBs.
We would have to wait for the camera to finish these 100TRBs before doing anything.

Queuing a stop endpoint command is essential to canceling a URB, and we can't
wait for the ring to be empty before issuing it.

-Mathias